a fuzzy logic based short term load forecast for …load forecasting but there are fewer studies on...
Post on 13-Mar-2020
4 Views
Preview:
TRANSCRIPT
Abstract—Electric load forecasting is important for economic
operation and planning. Holiday load consumptions are very
different than normal days and does not follow the regular trend
of normal days. Also data for the holidays less than other normal
days. So making an accuracy holiday load forecast model is a
difficult task. The purpose of this paper is presents different
models using fuzzy logic method without weather information.
Firstly holidays are classified according to their characteristics
and historical load shapes. Each fuzzy model have three inputs
and one output. While historical data from past years,
consumption data from last week and the type of holiday
(national, religious) are selected as inputs, the output is hourly
forecasted holiday load. The data between the years 2009 and
2011 are used to design the forecasting models. The model
performances are evaluated with the real data of the year 2012.
The results of models are compared each other and show that
proposed Model 2(scaled model) is more successful than Model
1. This paper shows that fuzzy logic can give good results for the
holiday short term load forecast.
Index Terms—Fuzzy logic, holiday load forecasting,
short-term load forecasting.
I. INTRODUCTION
The electricity used by the consumer should have some
certain characteristics which are continuous, safe, reliable and
minimum price. To perform all these characteristics, variety
of plans should be prepared for power system. Electrical load
forecasting is the first step of this plans and it is divided into
four types: long term, medium term, short term and very short
term. The short term load forecasting is a prediction of load
from 1 hour to 1 week and provides these characteristics by
helping the energy system operators to make efficient energy
management operations and better power system planning.
These operations and plans include energy purchasing, unit
commitment; reduce spinning reserve capacity and T&D
(transmission and distribution) operations [1]-[4].
The forecasting accuracy is very important factor for power
system efficiency. If the forecast is overestimated, it leads to
the start-up of too many units supplying an unnecessary level
of reserve. Thus the production cost of is increased. In
addition to that, it leads to substantial wasted investment in
the construction of excess power systems. On the contrary if
the forecast is underestimated, it may result in a risky
operation and unmet demand persuading insufficient
preparation of spinning reserve. Besides it may cause the
system to operate in a vulnerable region to the disturbance. As
Manuscript received October 5, 2015; revised January 11, 2016.
The authors are with the Selçuk University, Technology Faculty,
Department of Electrical & Electronics Engineering, Konya, Turkey
(e-mail: hasanhcevik@selcuk.edu.tr, mcunkas@selcuk.edu.tr).
a results, the frequency can drop and power outages can occur
at power systems [5], [6].
The energy management system wants the hourly load
consumption forecasts for next day from the distribution
companies and the major consumer such as iron and steel
factories one day in advance. This system, called day ahead
program, helps to perform the following goals
provide market participants the opportunity balance the
generations and consumption
provide the system operators a balanced system in the day
ahead
determine reference price for electric energy
provide market participants some opportunities for the
following day [7].
The trends of load consumption change according to the
time. For example the weekdays load profiles and the
weekends load profiles are different from each other, but
holidays load profiles are very different than everything else.
The electricity consumption decreases excessive amounts in
holidays, therefore new forecasting models should be
developed for holidays. In addition to these, holiday load
forecasts are harder than normal days because of small
number of available historical data for them compared to
other days [8].
There are many studies under the umbrella of short term
load forecasting but there are fewer studies on holiday load
forecast. Holidays are also referred to as abnormal days or
special days in some studies in the literature. A forecast for
special days is carried out using artificial neural network
(ANN) and fuzzy inference method [9]. While 24 hourly
scaled load values of the same special day in the previous year
and information of the special day type are selected for inputs
ANN, the 24 hourly scaled load values are the outputs. In
addition to that, the fuzzy logic is used for minimum and
maximum load forecasting procedure. The forecasting results
are obtained with combination of the two systems results. In
another study, an hourly load forecast of the anomalous days
is performed by means of self-organising map and ANN [10].
In addition there are some holiday forecasting studies using
fuzzy linear regression method [11], fuzzy inference system
[12], neural and fuzzy networks [13], Bayesian neural
network [14], linear regression [15], support vector machines
[16], fuzzy polynomial regression [8], semi-parametric
additive model [5].
This paper focuses holiday load forecast of Turkey using
different fuzzy models without weather information. Firstly
holidays are classified according to their characteristics and
historical load shapes. Previous year holiday data, previous
load and the type of holiday (national, religious) are selected
as inputs. The remainder of this paper is organized as follows.
A Fuzzy Logic Based Short Term Load Forecast for the
Holidays
Hasan H. Çevik and Mehmet Çunkaş
International Journal of Machine Learning and Computing, Vol. 6, No. 1, February 2016
57doi: 10.18178/ijmlc.2016.6.1.572
TABLE I: SOME INFORMATION OF HOLIDAYS IN YEARS 2009 TO 2012
Type Name Name 2009 2010 2011 2012 N
ati
on
al
New Year's Day 1 January Thursday Friday Saturday Sunday
National Sovereignty and
Children's Day 23 April Thursday Friday Saturday Monday
Labor Day 1 May Friday Saturday Sunday Tuesday
Commemoration of
Atatürk, Youth and
Sports Day
19 May Tuesday Wednesday Thursday Saturday
Victory Day 30 August Sunday Monday Tuesday Thursday
Republic Day 29 October Thursday Thursday Saturday Monday
Reli
gio
us Ramadan Feast Moves
19-22 September
Saturday-Tuesday
9-11 September
Thursday-Saturday
29 Aug-01 Sept
Monday-Thursday
18-21 August
Saturday-Tuesday
Sacrifice Feast Moves 26-30 November
Thursday-Monday
15-19 November
Monday-Friday
5-9 November
Saturday-Wednesday
24-28 October
Wednesday-Sunday
In Section II, holiday load profiles are analysed. While
Section III presents the forecast model, the results are given in
Section IV. Finally, the conclusions are in Section V.
II. HOLIDAY LOAD CHARACTERISTICS
The hourly holiday load consumption data are selected to
create the model for the years 2009 through 2011. The load
prediction is carried out for 2012 holidays. The load data are
obtained from a utility company from Turkish Electricity
Transmission Company (TEIAS), Turkey. As mentioned in
the previous section, since the load trends of holidays are
different, it is important to analyze the load data. Load data
should be analyzed before developing a prediction model for
achieving good forecast results. Public holidays vary from
country to country. In Turkey, there are two kinds of holidays,
national and religious. While national holiday dates do not
change, religious holiday dates change each year. The
holidays are presented in Table I and it can be seen some
holiday load curves in Fig. 1 and Fig. 2. The following
conclusions which help us to design forecast models can be
made by the analysis of Table I and Figures:
Each type of holidays’ load consumptions are less than
neighbor weeks’
National and religious holidays have different
characteristics. The reduction in load consumption is
higher for religious holidays.
Although some workplaces in private sector work on
national holidays, they do not work on Sundays. So, if the
national holiday is on Sunday, load consumption is
reduced more than the other national holidays which are
not on Sunday.
A day can be both on national holiday and on religious
holiday, because shifting in religious holidays each year.
For example 30 August 2011. In such cases, the day can
be considered as a religious holidays. The reason of this,
reduction in load consumption is higher for religious
holidays.
A national holiday can come after a religious holiday and
these effects the load curve of national holiday, for
example, 29 October 2012. In such cases, the national
holiday can be considered as a religious holidays.
If four weekdays are on holiday, the government can
officially announce that the whole week is holiday.
There are several factors which affect the load such as
historical load, time factors (day type, holiday type), weather
factor and random disturbances. We only use historical data
and time factors. We did not consider weather factor for this
forecast. This study presents two different fuzzy systems.
Holidays are grouped with load characteristics and so four
holiday types are determined.
Fig. 1. Some national holiday load curves of 2012.
Fig. 2. Ramadan feast holiday load curve of 2012.
Fig. 3. Inputs/output of fuzzy models.
National holiday
National Sunday holiday
Preparation day of religious holiday
Religious holiday
III. FUZZY LOGIC MODELS
Fuzzy logic is a powerful method with successive results
and it has been studied in many power system problems such
International Journal of Machine Learning and Computing, Vol. 6, No. 1, February 2016
58
as load forecasting, electricity price forecasting and power
system planning.
Fig. 4. Membership functions of input/output for Model 2.
In this study two different fuzzy logic models are
developed and tested for each holiday types using Matlab
Fuzzy Logic Toolbox. Each fuzzy model having three inputs
and one output are shown in Fig. 3. While inputs of Model 1
are previous year holiday load, previous week load and
holiday type, the inputs of Model 2 are scaled previous year
holiday load, previous week load and holiday type. The
output, hourly forecasted holiday load, is same for the models.
Also the previous year load data, previous week data and
forecasted load are normalized in both models. Each models
have different membership functions and rule bases.
As mentioned in previous section, there are some
similarities in terms of load decrease and increase time
although each type of holiday load consumptions are less than
neighbor weeks’.
It looks like they have only level difference between them
especially for national holidays. For this reason previous
week data from holiday can be scaled for using these
similarities. We developed a scaled factor for each forecast
hours on the previous week data before using the data for an
input for Model 2. Therefore we had the proportional
information from past year data and applied this information
to the data of this year. Scaled factor was calculated by using
Eq.(1).
N
N
NN PP
HPSP *1
1
11
(1)
where, P is previous week hourly load data, H is holiday
hourly load data, N is forecast year and SP is scaled previous
week hourly load data.
The membership functions input/output are shown in Fig. 4
for Model 2. The scaled previous year holiday load input has
five membership functions. Where VL is very low, L is low, N
is normal, H is high and VH is very high. VL and VH
membership functions are the generalized bell function and
the others are Gaussian functions. The membership functions
of the previous week load have three memberships (Fig. 6(b)).
These are L=low, N=normal and H=high. L and H
memberships are the generalized bell function and N
membership is Gaussian function. Holiday type has four
membership functions as mentioned in previous section and
their memberships are trapezoidal.
TABLE II: HOURLY FORECASTING RESULTS OF 1 JANUARY AND 23 APRIL
1 January Sunday 23 April Monday
Hours Actual
[MW]
Model 1
[MW]
Model 2
[MW]
Model 1
Error(APE)
Model 2
Error(APE)
Actual
[MW]
Model 1
[MW]
Model 2
[MW]
Model 1
Error(APE)
Model 2
Error(APE)
1 23312 22372 24592 4,03 5,49 21112 19525 19872 7,52 5,87
2 22168 20173 21911 9,00 1,16 19710 18910 18753 4,06 4,85
3 20815 19212 19172 7,70 7,89 18946 18869 18753 0,41 1,02
4 19904 18977 18753 4,66 5,78 18545 18857 18753 1,68 1,12
5 19327 18898 18753 2,22 2,97 18456 18861 18753 2,20 1,61
6 18940 18867 18753 0,39 0,99 18310 18865 18753 3,03 2,42
7 19066 18883 18753 0,96 1,64 17825 18872 18753 5,88 5,21
8 18416 18817 18753 2,18 1,83 19323 19027 18753 1,53 2,95
9 18997 18975 18753 0,11 1,28 22658 23316 23078 2,91 1,86
10 20376 20006 20154 1,82 1,09 24653 25248 24140 2,41 2,08
11 22138 22349 22491 0,95 1,59 25367 25564 24827 0,78 2,13
12 23644 24070 24054 1,80 1,73 25970 25886 25683 0,32 1,11
13 24638 23846 24598 3,22 0,16 25115 25293 24613 0,71 2,00
14 24965 24179 24653 3,15 1,25 25413 25372 24809 0,16 2,38
15 25143 23953 24665 4,73 1,90 25470 25449 25239 0,08 0,91
16 25316 23882 24727 5,66 2,33 25120 25250 24596 0,52 2,09
17 26516 26853 25347 1,27 4,41 24794 25157 24759 1,46 0,14
18 27889 27978 25780 0,32 7,56 24117 24416 23947 1,24 0,71
19 27919 27894 25779 0,09 7,66 23866 24026 23729 0,67 0,57
20 27784 27730 25779 0,19 7,21 24848 24731 24211 0,47 2,56
21 27317 27558 25779 0,88 5,63 26306 25284 25455 3,88 3,24
22 26836 26994 25780 0,59 3,94 25932 24905 24642 3,96 4,98
23 26392 27006 25780 2,33 2,32 25782 24862 24634 3,57 4,45
24 25353 24589 25336 3,01 0,07 24303 24044 23789 1,07 2,12
International Journal of Machine Learning and Computing, Vol. 6, No. 1, February 2016
59
IV. RESULTS
This study proposes two different fuzzy logic models for
holiday sort term load forecasting using real data. Four
different holiday types have been determined according to do
load curve rhythms. The forecasted results are compared with
the actual load. The fuzzy models and forecast accuracy
should be measured using some statistical functions. Mean
absolute percentage error (MAPE) widely used in load
forecasting studies and is selected to validate the proposed
models as follows:
100*1
1
n
t t
tt
R
FR
nMAPE (2)
where, R is actual value, F is the forecasting value and n value
is the number of forecasting.
Fig. 5. Actual and forecasted values for some national holidays.
Fig. 6. Actual and forecasted values for Ramadan holiday.
TABLE III: MAPE VALUES OF MODEL 1 AND MODEL 2 FOR 2012 HOLIDAYS
Type Name Date MAPE
Model 1 Model 2
Na
tio
na
l
New Year's Day 1 January Sunday 2,55 3,25
National
Sovereignty and
Children's Day
23 April Monday 2,10 2,43
Labor Day 1 May Tuesday 2,03 2,32
Commemoration of
Atatürk, Youth and
Sports Day
19 May Saturday 3,68 5,58
Victory Day 30 August Thursday 9,03 9,38
Republic Day 29 October Monday 11,29 7,16
Reli
gio
us
Preparation day of
Ramadan Feast 18 August Saturday 4,59 4,57
Ramadan Feast 19-21 August
Sunday-Tuesday 7,54 5,24
Preparation day of
Sacrifice Feast
24 October
Wednesday 7,04 5,13
Sacrifice Feast 25-28 October
Thursday-Sunday 4,66 6,16
Weighted
Average 5,57 5,34
Load forecasting was carried out totally for 13 days (312
hours). Some forecasting results are presented in Fig. 5, Fig. 6
and Table II. As can be seen from the figures, trends of actual
load curve and forecasted load curves are similar. The
national holidays’ consumptions were forecasted well than
the Ramadan holiday consumptions in each model.
The comparison of 2012 holiday results for Model 1 and
Model 2 are shown in Table III. Maximum MAPE value and
minimum MAPE value are obtained as 11.29 and 2.03,
respectively, for all holidays. Also the weighted MAPE
values are obtained as 5.57 and 5.34, respectively, for Model
1 and Model 2. We can say that while Model 1 has been more
successful on national holidays, Model 2 has been more
successful on religious holidays.
V. CONCLUSION
In this study, load forecasting models for holidays is
developed by using fuzzy logic. Holidays are grouped with
load characteristics. Two hourly holiday load forecast models
are developed and the results of models are compared. We did
not consider weather factor for this forecast. Inputs of Model
1 are previous year holiday load, previous week load and
holiday type. The previous year holiday load input is scaled
for Model 2 unlike Model 1. The sample averages of MAPEs
for Model 1 and Model 2 are 5.57 % and 5.34 %, respectively
for all holidays of 2012. It can be said that both models find
close values to the real data and produce good forecasting
results. The results show that Model 2 could have better
forecasting accuracy than Model 1 in terms of MAPE values.
ACKNOWLEDGMENT
This study has been supported by Scientific Research
Project of Selcuk University.
REFERENCES
[1] M. Q. Raza and A. Khosravi, "A review on artificial intelligence based
load demand forecasting techniques for smart grid and buildings,"
Renewable and Sustainable Energy Reviews, vol. 50, pp. 1352-1372,
2015.
[2] T. Hong, P. Pinson, and S. Fan, "Global energy forecasting competition
2012," International Journal of Forecasting, vol. 30, pp. 357-363,
2014.
[3] H. H. Cevik and M. Cunkas, "A comparative study of artificial neural
network and ANFIS for short term load forecasting," in Proc. 6th
International Conference on Electronics, Computers and Artificial
Intelligence, 2014, vol. 5, pp. 29-34.
[4] H. H. Çevik and M. Çunkaş, "Short-term load forecasting using fuzzy
logic and ANFIS," Neural Computing and Applications, vol. 26, pp.
1-13, 2015.
[5] S. Fan and R. J. Hyndman, "Short-term load forecasting based on a
semi-parametric additive model," IEEE Transactions on Power
Systems, vol. 27, pp. 134-141, 2012.
[6] H. H. Çevik, "Short term electrical load forecasting of Turkey," M.S.
thesis, Selçuk University Directorate of Graduate School Natural and
Applied of Science, Konya-Turkey, 2013.
[7] Electricity market balancing and settlement regulation, Offical Gazette
of the Republic of Turkey No.27200, 2009.
[8] Y.-M. Wi, S.-K. Joo, and K.-B. Song, "Holiday load forecasting using
fuzzy polynomial regression with weather feature selection and
adjustment," IEEE Transactions on Power Systems, vol. 27, pp.
596-603, 2012.
[9] K.-H. Kim, H.-S. Youn, and Y.-C. Kang, "Short-term load forecasting
for special days in anomalous load conditions using neural networks
and fuzzy inference method," IEEE Transactions on Power Systems,
vol. 15, pp. 559-565, 2000.
International Journal of Machine Learning and Computing, Vol. 6, No. 1, February 2016
60
[10] G. Chicco, R. Napoli, and F. Piglione, "Load pattern clustering for
short-term load forecasting of anomalous days," in Proc. IEEE Porto
Power Tech Proceedings, 2001, vol. 2, pp. 6.
[11] K.-B. Song, Y.-S. Baek, D. H. Hong, and G. Jang, "Short-term load
forecasting for the holidays using fuzzy linear regression method,"
IEEE Transactions on Power Systems, vol. 20, pp. 96-101, 2005.
[12] Q. Ding, H. Zhang, T. Huang, and J. Zhang, "A holiday short term load
forecasting considering weather information," in Proc. The 7th
International Power Engineering Conference, 2005, pp. 1-61.
[13] S. Nahi, "Load forecasting on special days & holidays in power
distribution substation using neural & fuzzy networks," in Proc.
International Conference on Computational Intelligence for
Modelling, Control and Automation, 2006, pp. 118-123.
[14] N. Mahdavi, M. Menhaj, and S. Barghinia, "Short-term load
forecasting for special days using bayesian neural networks," in Proc.
Power Systems Conference and Exposition, 2006, pp. 1518-1522.
[15] R. S. Fernandes, Y. K. Bichpuriya, M. Rao, and S. Soman, "Day ahead
load forecasting models for holidays in Indian context," in Proc.
International Conference on Power and Energy Systems, 2011, pp.
1-5.
[16] Y. Cai, Q. Xie, C. Wang, and F. Lü, "Short-term load forecasting for
city holidays based on genetic support vector machines," in Proc. 2011
International Conference on Electrical and Control Engineering,
2011, pp. 3144-3147.
Hasan Hüseyin Çevik was born in Konya, Turkey
in 1984. He received his BS and MS degree from
the Department of Electrical-Electronics
Engineering, Selçuk University in 2009 and 2013
respectively. Since 2011, he is working as a
research assistant in the Technology Faculty
Department of Electrical & Electronics, Selçuk
University. His main research interests are in the
area of artificial applications and their applications
in power systems.
Mehmet Çunkas received the BS, MS, and Ph.D
degrees from the University of Selcuk, Turkey, all
in electrical engineering in 1994, 1998, and 2004,
respectively. He has had extensive experience in the
field of electrical machines, both in industry and as
an academic. He is currently working as Assoc.
Prof. at the Electrical Engineering Department,
Technology Faculty, University of Selcuk. His
interests include dynamic modeling of electrical
machines using numerical methods, design and analysis of novel
electromechanical energy conversion devices, artificial intelligent
techniques, intelligent optimization algorithms, and application of
mathematical techniques to design optimization of electromagnetic
devices.
International Journal of Machine Learning and Computing, Vol. 6, No. 1, February 2016
61
top related